AI news: June 2026

June 2026 ended with a packed June 30: Anthropic shipped Claude Sonnet 5 and Claude Science on the same day the US lifted export controls on Fable 5 and Mythos 5, and Google released Nano Banana 2 Lite. Days earlier, OpenAI had opened the GPT-5.6 Sol preview and Meituan introduced LongCat-2.0, the first giant Chinese model trained entirely on domestic chips.

Claude Sonnet 5 lands in the API

Anthropic launched Claude Sonnet 5 on June 30, generally available across all plans and the new default on Free and Pro. Introductory pricing through August 31: $2 per million input tokens and $10 per million output (then $3 and $15, same as Sonnet 4.6). It brings a 1-million-token context window, up to 128k output and adaptive thinking on by default. Anthropic cites 81.3% on OSWorld-Verified; Cursor reports 57% on its CursorBench, up from Sonnet 4.6’s 49%. One footnote from the docs: the new tokenizer produces roughly 30% more tokens for the same text, so per-task cost can shift even though the rate doesn’t.

For most production workloads this is the natural upgrade: more capability at the same nominal price. Measure the tokenizer effect on your bills before booking the savings; among practitioners, the reported cost edge over Opus 4.8 holds at medium effort and thins out at high effort.

Source: Anthropic · also on TestingCatalog and The Deep View

GPT-5.6 Sol, previewed for about twenty partners

OpenAI introduced the GPT-5.6 family on June 26: Sol (flagship, $5 per million input and $30 output), Terra ($2.50 and $15) and Luna ($1 and $6). Sol adds an “ultra mode” with subagents; technical press reports 88.8% on Terminal-Bench 2.1 (91.9% in ultra), ahead of GPT-5.5 and Claude Mythos 5 on that test. Access is a preview limited to some 20 organizations chosen by direct contact, with no public waitlist and no ChatGPT presence; OpenAI shared the model and its rollout plan with the US government before widening access, and reports over 700,000 GPU-hours of red teaming.

There is nothing to adopt yet: this is the pricing and tier announcement ahead of the next family’s rollout. The gated release with prior government review arrived in the same month as the Fable 5 episode.

Source: TestingCatalog · also on The Deep View and Neat Prompts

Claude Science: Anthropic’s workbench for scientists

Also on June 30, Anthropic released Claude Science in public beta for paid plans: a coordinating agent with over 60 skills and preconfigured connectors for genomics, proteomics, structural biology and cheminformatics, access to databases like UniProt, PDB, Ensembl and ChEMBL, and integration of the NVIDIA BioNeMo Agent Toolkit with the Evo 2, Boltz-2 and OpenFold3 models. It runs compute on a laptop, on HPC clusters over SSH, or on on-demand GPUs via Modal. The “AI for Science” program funds up to 50 projects with credits of up to $30,000 each.

It is the Claude Code skills play applied to the lab: a generalist agent plus the domain’s tool scaffolding. For research teams, the actionable part is that it runs on your own infrastructure, laptop to cluster.

Source: Anthropic · also on MIT Technology Review and The Deep View

LongCat-2.0: 1.6 trillion parameters with no Nvidia hardware

Meituan released LongCat-2.0 on June 30: a 1.6-trillion-parameter MoE (48B active), a 1-million-token context, pretrained on over 35 trillion tokens on a cluster of some 50,000 domestic chips, with no Nvidia in the process. Meituan says it is the largest Chinese model trained entirely on national hardware; several sources identify the chips as Huawei Ascend, without official confirmation. It cites 59.5 on SWE-bench Pro, ahead of GPT-5.5 per those same sources. The API is compatible with OpenAI and Anthropic formats ($0.75 per million input, $2.95 output); the weights, announced under MIT, were not yet published at launch.

The news is industrial more than benchmark-shaped: full pretraining off Nvidia hardware was the line left to cross. As a usable model, the API compatibility makes it cheap to trial, with the weights still pending.

Source: SCMP · also on TestingCatalog and Hugging Face

MiniMax M3: coding, 1M context and multimodality in one model

MiniMax opened the month on June 1 with M3: a 1-million-token context via its MSA sparse attention and, per the company, the first open-weight model combining frontier coding, that context and native image and video multimodality. The benchmarks (59.0% on SWE-Bench Pro, 83.5% on BrowseComp) are self-reported and externally unverified, as The Decoder notes. At launch the weights were not published: MiniMax promised them “within 10 days”, so “open-weight” was, that day, a promise. Secondary sources cite $0.30 per million input tokens.

The announced combination is exactly what cheap multimodal agents call for; until the weights are up and someone external replicates the numbers, it is a candidate to test through the API, not a decision.

Source: MiniMax · also on The Decoder and VentureBeat

Microsoft Build 2026: in-house models and the Copilot app

At Build (first week of June), Microsoft introduced MAI-Thinking-1, its first in-house reasoning model: a 35B MoE with a 256k context, trained without distilling from other providers, scoring 97.0% on AIME 2025 per Microsoft and preferred over Claude Sonnet 4.6 in blind evaluations by its rating partner Surge (vendor figures); it is in private preview on Microsoft Foundry. Alongside it came MAI-Code-1-Flash, a budget coding model rolling into the GitHub Copilot model picker in VS Code, and the GitHub Copilot App, a desktop app for steering several coding agents in parallel with a view of sessions, issues and pull requests.

The underlying move is Microsoft cutting dependence on third-party models in its own stack. For teams working on GitHub, the short-term tangibles are the Flash model in Copilot’s picker and the desktop app.

Source: TestingCatalog · also on Microsoft Foundry and Tech Times

DiffusionGemma: text by diffusion, 4x faster

Google DeepMind released DiffusionGemma on June 10: an experimental open model (Apache 2.0) that generates text through discrete diffusion instead of the usual token-by-token autoregression. It is a 26B MoE built on Gemma 4 with 3.8B active parameters, generates blocks of up to 256 tokens in parallel and, per Google, exceeds 1,000 tokens per second on an H100 and 700 on an RTX 5090; quantized, it fits in 18 GB of VRAM. Google itself states the trade-off: output quality below standard Gemma 4 on all its internal benchmarks, with in-line editing and code infilling as the target cases.

It is the first serious chance to try diffusion-based text generation locally: for autocomplete and interactive editing, where latency beats peak quality, the trade can pay off.

Source: Google · also on Hugging Face and The New Stack

Fable 5’s month: suspension, a new classifier and redeployment

Claude Fable 5 launched on June 9 with, per Anthropic, the strongest safeguards it has ever applied to a model (same base as Mythos 5, which stayed restricted to defensive-cybersecurity partners). Three days later, the US imposed export controls after a report from Amazon researchers: pressed to identify software vulnerabilities, in one case the model produced code demonstrating how to exploit one. Anthropic suspended global access for 19 days since it could not verify user nationality in real time, built a classifier that blocks the reported technique in over 99% of cases, published the CJS-0 to CJS-4 jailbreak-severity scale with Amazon, Microsoft and Google (Project Glasswing), and opened a HackerOne reporting program. Access was restored on July 1.

For anyone building on frontier models, the episode leaves two operational lessons: a model can vanish overnight by regulatory order, and model contingency plans stop being theoretical. Anthropic’s tests showing less capable models reproduced the same exploits lower the technical severity of the original finding.

Source: Anthropic · also on GBHackers and BleepingComputer

Nano Banana 2 Lite and Gemini Omni Flash, for builders

On June 30 Google shipped two developer-facing pieces. Nano Banana 2 Lite (generally available): image generation and editing at 4 seconds per image, about 5x faster than Nano Banana 2, at $0.034 per 1K image ($0.0168 in batch), with legible text and SynthID; external coverage notes weaker small text and character consistency than the full model. Gemini Omni Flash (preview): the API for the video model shown at I/O, with multi-turn conversational editing, clips capped at 10 seconds and pricing of $0.10 per second at 720p.

For product work, Lite puts image generation in the cost range where it slots into user flows without a second thought; Omni Flash lets you prototype conversation-editable video, with the duration limits clear from day one.

Source: Google · also on Ars Technica and eesel.ai

The US lifts export controls on Fable and Mythos

The month’s regulatory arc closed on June 30: the Commerce Department lifted the export controls imposed on June 12 on Fable 5 and Mythos 5. Per Commerce Secretary Howard Lutnick, Anthropic will no longer need an export license in exchange for committing to proactively detect security risks, collaborate on standards for future models and report malicious activity. Mythos 5 had already been restored for some US organizations on June 26; Fable 5 returned to global availability on July 1.

The outcome fixes the access framework for the most capable models: global availability for the general line (Fable) and approved-organization access for the unrestricted variant (Mythos).

Source: CNBC · also on Washington Post and Nextgov/FCW